Abstract

Metal additive manufacturing (AM), especially laser powder bed fusion (LPBF), enables the production of near-net-shape metallic components with complex internal features and part consolidation directly from digital designs, offering capabilities that are difficult or impossible to achieve with conventional manufacturing routes. LPBF is widely regarded as a cornerstone within digital manufacturing and the broader Industry 4.0 paradigm. However, despite its potential, LPBF still faces major challenges in producing defect-minimized structural components. Chief among these is the generation of high residual stresses, which can lead to excessive part distortion, crack initiation and propagation, and even structural failure. These issues arise from the highly localized and rapid heating and cooling cycles and associated phase transformations that occur during layer-by-layer melting and solidification process.
To address these challenges, the core objective of this research is the development of a high-fidelity, experimentally validated, multi-scale finite element (FE) modeling framework capable of accurately predicting residual stresses and distortions in LPBF-manufactured parts. A sequentially coupled thermo-mechanical simulation, based on the modified inherent strain (MIS) method, was implemented in Abaqus. In this framework, detailed thermo-mechanical simulations with a moving heat source and temperature-dependent material properties are first used to extract laser-induced thermal histories and applies the resulting thermal loads in the mechanical simulation to calculate modified inherent strains. These inherent strains are then fed into a reduced-order mechanical model to efficiently predict residual stresses and distortions. Several custom Python scripts, and user subroutines were developed for modeling phase transformations, hardening behavior, moving heat sources, and progressive element activation in Abaqus. To bridge the gap between microscale and macroscale modeling, a layer agglomeration technique was introduced, allowing many physical layers to be represented by fewer equivalent computational layers, thereby reducing simulation time without compromising accuracy. Model validation was conducted using Inconel-718 cantilever beam samples in which various plasticity models including elastic-perfectly plastic, isotropic hardening, kinematic hardening, and Johnson-Cook, were evaluated and compared. The kinematic hardening model showed the best agreement with experimental results, with less than 6% deviation and significantly improved computational efficiency. Additionally, the MIS model's capability to predict subsurface residual stresses was evaluated using cube coupons made of Inconel-625, addressing a critical gap in the existing literature.
To further accelerate the simulation process, a machine learning (ML) model was trained using data from the developed high-fidelity models to predict inherent strains directly from process parameters. The ML model could estimate the strains with an average relative error of 18% for three material types compared to FE results. When embedded in the FE framework, this hybrid ML–FE approach reduces computational cost by up to four orders of magnitude (⁓10,000×) while maintaining reliable accuracy, with residual stress and distortion predictions within 16% of the high-fidelity reference solutions.
Overall, this work provides a robust, scalable framework for predictive modeling and process optimization in LPBF, paving the way toward reliable, reduced-distortion AM production.